[CUDA] Relax restrictions on variadics in host-side compilation.

[tra]

Summary

D150718 allows variadics during GPU compilation, but we also need to do it for
the host compilation as well, as it will see the same code.

clang/lib/Driver/ToolChains/Clang.cpp

else {
         // determines how we load and launch GPU kernels.
         auto *CTC = static_cast<const toolchains::CudaToolChain *>(
             C.getSingleOffloadToolChain<Action::OFK_Cuda>());
         assert(CTC && "Expected valid CUDA Toolchain.");
         if (CTC && CTC->CudaInstallation.version() != CudaVersion::UNKNOWN)
           CmdArgs.push_back(Args.MakeArgString(
               Twine("-target-sdk-version=") +
               CudaVersionToString(CTC->CudaInstallation.version())));
+        // Unsized function arguments used for variadics were introduced in
+        // CUDA-9.0. We still do not support generating code that actually uses
+        // variadic arguments yet, but we do need to allow parsing them as
+        // recent CUDA headers rely on that.
+        // https://github.com/llvm/llvm-project/issues/58410
+        if (CTC->CudaInstallation.version() >= CudaVersion::CUDA_90)
+          CmdArgs.push_back("-fcuda-allow-variadic-functions");
       }
     }
     CmdArgs.push_back("-aux-triple");
     CmdArgs.push_back(Args.MakeArgString(NormalizedTriple));
   }
 
   if (Args.hasFlag(options::OPT_fsycl, options::OPT_fno_sycl, false)) {
     CmdArgs.push_back("-fsycl-is-device");